Agents to the Rescue?
PATRICIA M. WEST Ohio State University DAN ARIELY
Massachusetts Institute of Technology STEVE BELLMAN
University of New South Wales ERIC BRADLOW University of Pennsylvania JOEL HUBER Duke University ERIC JOHNSON University of Pennsylvania BARBARA KAHN University of Pennsylvania JOHN LITTLE
Massachusetts Institute of Technology DAVID SCHKADE
University of Texas, Austin
Abstract
The advent of electronic environments is bound to have profound effects on consumer decision making. While the exact nature of these in¯uences is only partially known it is clear that consumers could bene®t from properly designed electronic agents that know individual users' preferences and can act on their behalf. An examination of the various roles agents perform is presented as a framework for thinking about the design of electronic agents. In addition, a set of goals is established that include both outcome-based measures, such as improving decision quality, as well as process measures like increasing satisfaction and developing trust.
Key words: agents, e-commerce, consumer choice
Introduction
The world of commerce is changing. Due to the pervasive adoption of information technology, we are witnessing the evolution of a new marketplace. Unlike traditional
face-to-face retail settings where a good can be touched, seen, and even tasted, transactions are occurring in computer media environments which do not offer the opportunity for directly experiencing a product. This paper focuses on two aspects of these new environments: the ®rst is the design of such environments themselves. We ask how these environments will affect customer decision making. Our second focus is on how to help consumers make better decisions in these new environments. Here we focus on the concept of an electronic agent, and consider the roles they can play in computer-mediated environments.
Our interest in electronic consumer environments is motivated by two observations. First, in many ways these environments are importantly different from normal ways of conducting commerce. Transactions can occur more quickly and across greater distances than with traditional retailers. Digital goods, such as a digitized recorded performance, or a computer program can be discovered, tested, evaluated, purchased, and delivered all in a few minutes time, and these transactions can easily occur across borders. These environ-ments can also change very quickly and inexpensively. While it is impossible for physical stores to differentially arrange their shelves for each consumer, detailed customization is intimately possible in electronic markets. Finally, these environments can collect and use abundant data about individual customers. In principle, and to some extent in practice, electronic environments allow us to not only observe what is purchased, but also what information is examined on the way to purchase.
However, one observation is unmistakable: despite the increase in computing speed touted by Moore's law, a particular CPU has not changed its capacity: that of the human decision maker. Ultimately, although this decision maker may be faced with more information, a wider variety alternatives, and other bene®ts, the processing of this information occurs using the same limited information processing capacity as in physical shopping environments. This constraint, unaffected by new technologies, raises an obvious question of; how these new environments affect consumer decisions.
To examine these questions, we look at two aspects of these new decision environments. First we ask how the environment itself may affect consumers decisions, and then we look at the prospects of electronic agents aiding these decision makers.
How Do Electronic Environments Affect Decision Makers?
A casual examination of the popular press suggests that the rise of electronic commerce will be accompanied by an increase of consumer sovereignty. Because the cost of search is reduced, consumers should be empowered and able to search for better products at lower prices (Bakos, 1997). The emerging view is one in which many marketplace imperfections disappear and one in which consumers are in a more powerful position compared with retailers. Alba et al. (1998; Lynch and Ariely, 1998) suggest that this fear is a major reason why many retailers are hesitant to move to the arena of electronic shopping.
However, we think the story may be a bit more complicated. Consumer choice, whether in a physical or electronic environment seems increasingly to be jointly determined by both a consumer's preferences and the features of the task environment. More recent behavioral views suggest several ways in which the choice environment might affect what a consumer
chooses. These views are particularly relevant to electronic environments because the design of these environments is so easy to change. In the next section we review some of this literature.
Preference Construction and Discovery
Many traditional models of consumer choice assume that a consumer's tastes are well articulated, and much like psychophysical functions. A more recent, evolving view suggests that for some kinds of preferences, consumers are constructing guesses about what they prefer (Bettman, Luce, Payne, 1998; Payne, Bettman, and Johnson, 1993; Fischhoff, 1991). Not only are such statements of preferences often constructed on the spot, but they represent at best guesses about what would really maximize hedonic pleasure (Kahneman and Snell, 1990; Lowenstein and Schkade, 1999 review this literature). These guesses are likely to be affected in online environments due to their ability to manipulate the context in which the choice is made.
One example is research by Mandel and Johnson (1998) who show that wallpaper, the background of a web site, can prime the importance of product attributes, and thereby alter choices even with a single brief exposure. Another example comes from Menon and Kahn (1998) who illustrate a possible connection between a users' initial experience with a Web site and subsequent choices, (e.g., a higher level of initial stimulation produces lower tendencies to choose novel products, less responsiveness to promotional incentives, and fewer unplanned purchases). The ¯exibility of the electronic environment raises both a caution and an opportunity. If electronic retailers can alter environments to bias choices in their favor, social welfare could drop. However, opportunity arises as increased competi-tion among electronic retailers make it less likely that retailers will survive who convince customers to purchase products that are not in their best interest. The relative in¯uence of these two factors will be important issues for further research.
Information Search and Organization
A second element important for online environments concerns the organization of information. For years, we have known that the organization of product information can in¯uence choice. The standard rationale is that the organization of information can change the cost of searching for various types of information, which in turn can in¯uence decision strategies (Bettman, Johnson, and Payne, 1990; Kleinmuntz and Schkade, 1993). Russo (1977) for example showed that reorganizing unit price information into lists that are sorted from cheapest to most expensive changed consumer choices in supermarkets. Can this happen in the virtual shopping environment?
Lynch and Ariely (1998) created online wine stores that manipulated the processing cost of three kinds of information: (1) price, (2) quality, and (3) the ability to compare between different online stores. They found that price sensitivity was highest when price was easy to process, but quality information was dif®cult to process. In contrast, when quality
information was easily generated, this attribute grew in importance resulting in reduced price sensitivity. Moreover, when actually given the wines for later consumption, subjects reported that they were happier with the wines purchased in the environment with easy access to quality information. Lohse and Johnson (1998) report similar results examining choice of disposable cameras and cookies.
The revolutionary power of the online environments is that the display of information is very malleable, and under the control of the seller, buyer, or both. Peapod, for example, allows customers to sort for themselves on a variety of attributes. A health conscious consumer can sort products based on ®ber or sugar content, for example, and a bargain hunter can sort on unit price. In contrast, we could easily imagine a retailer who allows sorting only on attributes for which they excel or those that give them a competitive advantage. From a research perspective, it is even more important to understand the impact of environmental contexts on product search.
Product Evaluation
Many product categories are not easily decomposed into constituent, Lancasterian features. For example, choices among movies, novels, ®ne art, and perhaps even homes, do not seem satisfactorily described by a small set of features common to all alternatives. When faced with these decisions, consumers often turn to others for advice. Consumers listen to critic opinions when deciding which movie to go see, discuss the books they read with friends and co-workers, and rely on an agent to ®nd a new home. In other cases, the number of alternatives that are available is immense, and with electronic environments, we suspect growing. For example, Amazon (www.amazon.com) claims to have available over 3 million titles, and CDNow (www.cdnow.com) carries over 300,000 CDs, cassettes, vinyl albums, music videos, laserdiscs, DVDs, movies and T-shirtsÐten times the size of the average music store.
The problem faced by consumers in these environments is serious. On the one hand, for decomposable products, search costs can lead to an increased reliance on satis®cing instead of optimization. Or these high variety product sets can cause consumers to delay purchase, not only because the complexity of the choice set is high, but also because consumers are uncertain as to the set of possible options (Greenleaf and Lehmann 1995). For non-decomposable products, the situation is more serious to the extent that an online environment can potentially rob one of important experiential information. Such environ-ments limit one's ability to ask a trusted friend her opinion of a movie, or to ask a store clerk how a novel compares to the author's prior work. Currently, E-bay (www.ebay.com) solicits users to provide self-reports of experiences with vendors as a means of building credibility and trust. Numerous movie and restaurant review sites, as well as online booksellers use similar self-sustaining mechanisms. To the extent that people ®nd layman reviews informative and credible they may turn to these alternative sources of input instead of agent services. However, it is not clear how a user can directly establish the credibility of this information. Witness, for example, the recent controversy over Amazon's decision to put ``preferred advertisers'' and the top of its recommendation lists. The growth in virtual
communities and use of collaborative ®ltering techniques to help identify like-minded users may be a direct response to consumers' need for experiential information and advice. Thus, electronic environments can be expected to replace small numbers of personal reviews with pooled recommendations from a large number of like-minded consumers.
Thus a major consideration in developing good online shopping environments is to provide tools to help solve these problems. We suggest that help is needed on at least three fronts: First, people need assistance in preference construction and discovery. Second, they need help ®nding and organizing relevant information. Finally, help is needed in evaluating attractive alternatives and executing decision strategies.
Agents to the Rescue?
The traditional de®nition of an electronic agent is a software program that ``knows'' users' preferences and can act autonomously on their behalf (Maes 1995). These systems typically require consumers to engage in an interactive process that involves sharing information, performing tasks consistent with the user's goals, and monitoring responses in order to learn how to improve performance. Given our concerns about consumer decision making in online environments we will examine how electronic agents might address some of these problems.
Of course, agents are not a new development (Solomon 1986). Consumers have used human agents, such as realtors, ®nancial analysts, and interior decorators, to assist them in their buying decisions or to act on their behalf long before the advent of the electronic marketplace. Interestingly, Reeves and Nass (1998) have demonstrated that people tend to interact with computers in a manner similar to how they interact with other people. Therefore, by examining the various bene®ts that human agents offer we hope to shed light on the features that one would want to incorporate into electronic agents.
Roles Agents Perform
Human agents are turned to for their expertise and advice as well as trusted to act as surrogates who save us time and reduce cognitive effort. Similarly, the ideal electronic agent would serve a variety of functions (see Table 1). Agents can act as tutors who inform clients about the most important factors to consider when evaluating alternatives within a product category, and to assist their clients in discovering their preferences. For example, a good physician will not only treat her patients'symptoms, but also acquaint them with how to improve their diet and lifestyle to maintain good health.
Agents can act as clerks who save us time and effort by screening out unattractive alternatives. In addition, they can provide access to otherwise closed markets, and facilitate transactions between parties. For example, many interior decorators have direct access to merchants and manufacturers that are not readily available to consumers. Similarly, a realtor's job does not end once he has found his client the perfect house. The agent will
often help with many of the details required to close the sale and may be called on to help ®nd a good plumber or electrician.
One of the primary roles of a human agent is to act as an advisor and provide guidance to a client. For example, a ®nancial analyst is expected to offer sound advice regarding the appropriate portfolio mix to ensure that a client can achieve speci®c long-term ®nancial objectives. Sometimes agents are called upon to negotiate the terms of a purchase on a client's behalf, and thus take on the role of banker.
The various roles that an agent can perform all require building trust. The belief that another will act as an advocate and pursue what is in the best interest of a client is necessary for the success of agency. In the various arenas where professional agents are commonly relied on, laws have been instituted to protect the interests of the public and prevent these individuals from taking advantage of their clients. In fact, in certain arenas society has deemed that individuals must use an agent to act on their behalf. Only a doctor can prescribe drugs, and it is a lawyer's task to represent us in court. Looking forward, it is likely that the same kinds of laws and institutional norms that protect people from abuse from human agents will develop for electronic ones.
Goals of an Electronic Agent
Our examination of consumer decision making in electronic environments and our consideration of the various roles that human agents play serves as a useful starting point for establishing a set of goals for electronic agents. These goals are summarized in Table 2. We will brie¯y elaborate on each of them.
Improving Decision Quality
There is enormous potential for electronic agents to assist users in navigating electronic environments and improving the quality of the decisions we make. Our discussion will parallel the roles of human agents described in Table 1.
Table 1. Role of Agents in Consumer Decision Making Consumer Decision Roles of Agents Description Making Tasks
Preference Construction
and Discovery Tutor Agents may educate a client about the features available in thecategory and help uncover client preferences. Information and
Alternative Search Clerk Agents may assist clients in performing tedious tasks such asinformation search and product screening. Product Evaluation Advisor Agents may be called upon to express their expert opinion, or
to provide advice tailored to their client.
Purchase Banker Agents may negotiate on their clients' behalf and facilitate the purchase of products and services.
Agent as Tutor. What is the ultimate goal of an electronic agent? If the goal is to improve consumer's decision making then it is very important that the agent can assess an individual user's familiarity with the product category and respond appropriately. Given the evidence that preferences are often constructed on the spot rather than re¯ecting a well-formed utility function one can not take for granted that users are able to articulate their true preferences directly. The fact that preferences are largely constructed suggests that measuring them is best viewed as a form of architecture (building a set of useful or defensible preferences) than as a form of archeology (uncovering preferences that are already there) (Payne, Bettman, and Schkade, in press).
The job of eliciting a user's preferences in order to provide a product recommendation is also not a simple task. Consider the following thought experiment. Suppose a (perverse) benefactor has decided to give you a laptop computer. Rather than allowing you to choose a particular model, she requires you to select between various ways to project your values and then used a machine to select the item with the highest score given these values. You can select from among four preference elicitation mechanisms. A model built form: (1) your past computer choices in the marketplace, (2) a series of hypothetical choice sets, (3) a ratings-based conjoint task, or (4) your self-explicated ratings of the importance of product attributes.
Thinking about this task makes it clear that eliciting consumer preferences is not a simple task. The ®rst two options, which are both choice-based, are appropriate for a user who is quite familiar with, and has practiced making decisions in the product category. A good example of how these two preference elicitation methods can fail is the in-store movie recommendation system called Take10, which Blockbuster recently tested. This system provided customers with a list of ten video recommendations based on the individual's previous rental history. However, the system never collected data indicating what the customer actually thought of the movie after viewing and did not control for multiple family members using the same card. Not surprisingly, the system failed to
Table 2. Goals of an Electronic Agent
Goal Features
Improve decision Assist consumers in learning about the product category and preference discovery quality Reduce decision complexity
Be a clearinghouse for good information
Provide an appropriate consideration set for evaluation Adapt over time in response to new information and feedback Increase satisfaction The system should be ¯exible and responsive to a user's requests
Increase the match between consumers' taste and the products they get Simple to operate and user-friendly
Eliminate tedious work and allow the consumer to experience the ``fun'' Set realistic expectations
Develop trust The agent must work to satisfy and protect the users' best interests Provide reassurance the user is getting the ``right product'' for a reasonable price
achieve customer acceptance and was withdrawn. The critical lesson here is that past choices may be inappropriate to use to predict future desires, thus, particularly in categories where people are prone to err in predicting their own tastes, it is important to collect post-consumption feedback as well as choice.
There are, however, situations in which a choice-based elicitation method might be appropriate. For example, when commitment from the consumer is an important part of the desired outcome. This may be one of the reasons contemporary society gives the individual the right to choose a mate, and parents allow children to make their own choice of college. Making the choice increases the likelihood that the choice will last due to personal commitment.
The second two preference elicitation methods are both ratings-based. These are more appropriate for the inexperienced consumer, for a product category containing novel products (artwork), or for infrequently purchased durable products (cameras). Both of these ratings-based tasks require considerable effort on the part of the user, but offer the opportunity to learn about features of the product category as well as one's personal preferences. The trade-off between the collecting additional information so the system can learn and improve its performance and the increased workload for the user is a dif®cult one to resolve. Avery and Zeckhauser (1997) point out the welfare problems associated with encouraging individuals to share information for the betterment of a community of users. They predict that incentives will be needed to prevent free riders from bene®ting from the goodwill of others. Even if the user is the direct benefactor of the effort, it may be dif®cult to motivate individuals to participate in an exchange of information.
One of the earliest electronic recommendation agents was Agent Inc.'s ``Fire¯y'' site. This system offered users personalized recommendations for movies and music. Personal pro®les were built by having users provide evaluations of movies they had already seen, or songs they liked. The developers used a ``collaborative ®ltering'' technique to identify like-minded users within its online community. The system's recommendations were based on the input from like-minded users. Today, several companies have deployed collaborative ®ltering methods to provide personal recommendations for books (www.amazon.com and www.barnesandnoble.com), music (www.launch.com), web sites (my.yahoo.com), and software (my.ZDNet.com). While collaborative ®ltering offers a relatively painless way for users to communicate their preferences and an effective means for developing useful product recommendations it cannot provide users insight into the drivers of their preferences (Gershoff and West 1998). A consumer who is unaware of the why he likes or dislikes an alternative will remain dependent on the agent to provide him with good recommendations, or worse, may continue to make bad choices in the future. If, on the other hand, a consumer actively works with the agent in the determination of preferences, he not only learns things about himself, but that kind of collaboration encourages the customer's buy-in and ultimate satisfaction with the choice (Kahn and Huffman 1998).
The foregoing suggests that consumers may require additional guidance, particularly in complex product categories, to help them understand the important features to consider when evaluating product alternatives. Unlike human agents, electronic agents tend to be outcome-focused (interested in providing good recommendations or ®nding all the sites that match a user's query) rather than process-focused (aiding the user in understanding the
category and why a product ®ts the person's needs). Incorporating a tutorial component into an electronic agent that encourages users to articulate and understand their values would provide numerous bene®ts to the consumer (Keeney 1992). An interesting parallel can be made to the study of human learning where it has been shown that people have dif®culty learning relationships without the help of cognitive feedback that informs them about the structure of the environment (Balzer, Doherty, and O'Connor, 1989). An educated consumer is more likely to choose the best product, and is therefore less likely to experience dissatisfaction in the consumption experience (West, Brown and Hoch 1996).
Agent as Clerk. Facilitating preference discovery and accurately eliciting user preferences are important goals to consider when building an electronic agent system. However, in some instances users are looking more for the information that an agent can provide than advice. Search engines are good examples of this. These systems perform the tedious task of searching for information or alternatives that ®t within a client's speci®cations. However, a focus on simply providing relevant information may not lead to good decisions (Ackoff 1967). This is particularly true for consumers who are inexperienced with the product category (Chase and Simon 1973). Information-overload, even when the informa-tion is ``good,'' may lead to sub-optimal decisions (Malholtra 1982; Ariely 1998). As stated earlier, while technology allows for an unlimited amount of information to be made available, our cognitive capacities have not changed and humans can only process a limited amount of information effectively. This suggests the need for ®nding ways to condense information into a form that is easier to process. An alternative by attribute table, for example, can be parsed much more ef®ciently and ultimately cause less confusion to the customer than serially presented alternatives (Huffman and Kahn 1998).
Sometimes consumers want advice that goes beyond the information they explicitly asked for. In these cases the electronic agent needs not only to search for the information that was requested by the consumer (clerk) but also to incorporate an advice giving mechanism and information ®ltering. PersonaLogic (www.personalogic.com) is designed to assist buyers in ®nding the perfect automobile or making travel plans by querying the user about the product features sought, then ®ltering out ``unacceptable'' alternatives. Similarly, Realtor.com is designed to help a potential buyer to screen homes that are currently for sale. The approach that both these systems adopt is quite ef®cient for users who are familiar with the product category and have a clear idea of what they are looking for. A user interested in scanning Realtor.com's database of over 100,000 homes nation-wide is asked to de®ne the neighborhood(s) being considered, an upper limit on price, the minimum number of bedrooms and bathrooms acceptable, etc. The system sequentially presents alternatives that ®t the user's speci®cations. Past research, however, has shown that using attribute cut-offs to screen alternatives tends to result in inferior product decisions due to inadvertent product elimination (Widing and Talarzyk 1993). This issue is particularly troubling in an agent context where the user may never be exposed to truly preferred alternatives because the system has eliminated them from consideration. Information ®ltering problems can be addressed by using ``fuzzy'' models that allow for uncertainty in the reporting of attribute cut-offs and=or by using collaborative ®ltering
methods to incorporate information from like-minded users. The problem of false con®dence in the search method suggests that agents need to include a monitoring system that enables them to test the degree of optimality of the choices made.
Agent as Advisor. In designing a system to provide consumers with personalized recommendations there are at least two competing goals operating. One goal is to provide a user with alternatives that are likely to be most desirable. Another important goal of the system is to learn and update its model of a user's preferences. Typically, each presented item is chosen to maximize its marginal probability of selection. In fact, many search engines sort Internet sites and some report the estimated selection probabilities. An open issue that arises from this approach is ``What is the appropriate set of items to present?'' Consider an Amazon visitor searching for a new book. In this case, showing all books with highest estimated selection probability is most likely to achieve this goal, yet there are drawbacks from both the agent's and user's point of view. For the agent, what learning about the user can take place? If the user selects a book the agent already ``expected,'' then the agent's learning is minimal. Since the only shown books are those most likely to be picked, which are highly correlated with each other, there is little opportunity for the agent to be ``surprised,'' and therefore learn. For the user, there is little opportunity to broaden one's consideration set as she is shown only those items most consistent with past behavior. If we believe that people desire variety (McAlister and Pessemier 1982, Kahn 1995), and=or become satiated with certain types of choices, then an intentionally broadened alternative item selection mechanism may be bene®cial. Such issues have previously been considered in diverse areas such as educational testing, optimal item selection in Computerized Adaptive Tests (Wainer and Mislevy, 1990), and experimental conjoint design, Adaptive Conjoint Analysis (Green, Krieger, and Agarwal 1991). In both cases, selection algorithms are designed to optimize information from the agent's perspective; the education of the user is not considered. However, such theory would provide a good starting point for research in optimal presentation sets for agents.
Two options are promising: (1) the system could show the user a set of alternatives that is a blend of most preferred alternatives and those which will provide the most information for the system (how to select these is an open issue), or (2) the system could ask the user a priori if the purpose is a single item search or learning about new alternatives. The idea of querying the user about intent has the added bene®t of reducing the risk of misinterpreting behavior when the user is engaged in choosing a product for someone else. The social dimension of buying for someone else, as occurs in gift selection creates yet another challenge for interpreting behavior. Amazon added a ``Gift Recommender'' to aid its customers in this dif®cult task. A side bene®t of this system is that it prevents misinterpreting an unexpected purchase (one that does not ®t with a user's personal history) as a change in preference. Clearly, the optimal presentation set and interpretation of a user's response will depend on the current goal.
Improving consumer decision quality will require continuous monitoring of a user's preferences and responses over time. Consumer needs change in response to life events and situational factors, such as the birth of a child or a new job. In addition, consumer preferences evolve over time as more experience is gained in a product category. For
example, an inexperienced wine drinker will favor the sweetness of Riesling to the density and richness of white Burgundy. However, as the individual's tastes develop he will come to appreciate the layers of buttery and citrus notes of the white Burgundy. This poses a challenge from a modeling perspective. How does the system detect changes in a user's behavior? How should the model weight old versus new observations? The solution to the problem is likely to require that the model regularly test itself by either asking the user to evaluate some unfavorable alternatives, or recommending unacceptable alternatives.
Many open issues exist regarding the generation of inferences about users by agents. One can separate these concerns into two areas: data and modeling. Regarding data, we see the main issues as: (1) how to summarize potentially massive amounts of information to allow predictions in real time, (2) how and when to let ``old information'' expire thereby refreshing the system, and (3) when to actually ``stop'' collecting information as user uncertainty is minimal. Summarizing users inputs=choices by a small number of informative but only partially suf®cient statistics, which are updated by adding new information and dropping old (``creating a ®ngerprint'') has been used in other areas (such as fraud detection for the phone company) and may be an area to pursue.
When modeling data to generate predictions, there are a number of research opportu-nities. First, data collected from an individual may be of many forms; e.g. choice, survey, and experimental studies. This is especially true given the possibility of using electronic agents to query users in multiple modes, and or obtain individual information from other sources. Therefore, how do we build ``supermodels'' which allow for incorporating and combining of seemingly uncongenial information? Secondly, recent research in collabora-tive ®ltering (Gershoff and West 1998) has suggested that when modeling or predicting individual preference, the utilization of other user's responses may play a signi®cant role (i.e. add information) even after conditioning on the features of the products. So developing models that incorporate product features and community knowledge to determine predictions for an individual user looks promising. Thirdly, models which present alternatives to users, and record which ones are selected are in fact collecting one additional piece of information; those options not selected. That is, there is information in seemingly missing data (Little and Rubin 1987) in that those ratings unobserved are likely to be lower, as they were not selected. Current collaborative ®ltering techniques do not incorporate this information. Finally, as with the data storage issues, these models will need to be implementable in real time, which is a rather daunting task.
Agent as a Banker. After the selection has been made, the ®nal role of the electronic agent is to take care of the ®nal transaction. This part is concerned mainly with ®nancial transactions and encompasses within it two main functions. The ®rst function is to select the vendor, supply, and payment terms. The second function is to transfer payment and ®nalize the transaction. In order to accomplish the ®rst function an agent can search all the potential vendors, and negotiate with them for appropriate terms (see Pazgal and Vulcan 1998). An example for such a system is currently tested by British telecom, where an agent negotiates in real time with different long distance carriers for price and bandwidth and uses the best alternative for any particular call. Unlike the other aspects of agent roles we
discussed, this aspect concerns only agent-agent communication and no human-agent dialog.
Increasing Consumer Satisfaction
While one would expect that a natural outcome of improving consumers' decision quality is increased overall satisfaction, however, this may not be the case. When measuring consumer satisfaction it is useful to consider both satisfaction with the decision process as well as with the ®nal choice (Fitzsimmons 1996; Fitzsimmons, Greenleaf, and Lehmann 1997). Just as the human component of any service exchange will impact a consumer's evaluation and future buying behavior, the interface of an electronic agent will directly affect a user's satisfaction with the decision process. Agents can make a client feel good about the process even if a poor choice is ultimately made (Widing and Talarzyk 1993). This is particularly true when there is a substantial probabilistic component associated with the outcome. For example, a ®nancial analyst might recommend purchasing a stock that winds up being a bad investment. Because most consumers know that the market is very dif®cult to predict they may not hold the analyst accountable for the loss, if the decision process was managed appropriately.
Flexibility and Responsiveness. A variety of steps can be taken to increase consumer satisfaction with the process and ®nal choice. Electronic agents have the unique ability to allow for personalization. Allowing the user to control the system to meet her personal needs will increase personal satisfaction (Gilmore and Pine 1997). PointCast (www.point-cast.com) is an interesting example of this. This screensaver delivers up to the minute news stories, stock prices, weather forecasts, and sports scores round the clock. Users are allowed to specify the type of information they are most interested in and to personalize the interface. It is also important to minimize the start-up costs associated with using the agent. Microsoft's ``intelligent'' agent, the paper clip, was speci®cally designed to anticipate a user's needs and offer assistance to reduce learning costs. Further it adapts to the usage patterns of the user, providing training that is appropriate to the user's current expertise.
One of the primary bene®ts an electronic agent can offer is to reduce the amount of tedious work the user needs to perform in the process of ®nding the information or product of interest. This may be as simple as storing user information, such as name, address, etc. for future visits so that the person does not have to continually provide the same information. However, this raises concerns about how this information will be used. Today many sites use the information collected in the form of cookies to personalize ads and other information on the user's screen when the user revisits a site. The primary reasons for Microsoft's recent acquisition of Fire¯y was because of its privacy technology that allows users more control of their personal information on the Internet in anticipation of consumer backlash from collection of clickstream data.
Managing Expectations. Managing consumer expectations is also an important factor to consider in the design of electronic agents. Early on, when the system has little information
to work with, the recommendations it provides may not suit the user well. Thus the users might lose faith in and stop using the system unless this is clearly communicated up front. One method of dealing with this problem is to collect enough demographic and back-ground information to establish reasonable priors using collaborative ®ltering. MovieCritic (www.moviecritic.com) is a good example here. This agent will not recommend a movie until the user has rated at least twelve ®lms and answered a battery of questions. If the user understands the importance of sharing information to improve the system's performance then he is more likely to be forgiving of mistakes early on. In addition, to managing consumer expectations regarding the agent's performance, it is also important to set realistic expectations of the product. One of the potential downsides of working with an agent that educates the user and presents a set of alternatives consistent with the user's speci®cations is that it forces the person to make dif®cult and unpleasant tradeoffs that might have gone unnoticed otherwise.
Increasing returns to scale associated with dealing with a single agent over time tend to create switching costs for the user. The more I use an agent, the more data the agent has on my tastes and, therefore, the better that agent can act on my behalf. In order to use another agent I will have to train it about my preferences. This feature invests electronic agents with a powerful loyalty enhancing property that is likely to keep users returning to site-speci®c agents. It also suggests a counterargument against the contention of friction-free capitalism on the Web.
Developing Trust: Who Does the Agent Work For?
The ®nal and arguably the most important requirement of a successful agent is that it develop and maintain trust (Urban 1998). Electronic agents are not fundamentally different in needing trust from humans and institutions in other aspects of commerce. Consider companies involved in relationship marketing with their partners, technical sales forces engaged in consultative selling, or brand managers defending the credibility of a brand. In these cases, problems of electronic and human agents are very similar.
Three dimensions of trust arise particularly in the context of electronic agents. First, trust is needed so that the consumer is willing to give information to the agent, either in the form of access to resources (e.g., a credit card number), or in the form of personal information (e.g., magazine readership). Clearly, legal and infrastructure changes will have to develop to assure con®dence in this critical component of electronic commerce. A second trust dimension constitutes the belief that the agent acts in the best interest of the customer. This trust is not absolute, but relative to expectations. For example, a reasonable expectation of a retailer's site (such as Amazon) is that it will ®nd the best books regardless of their pro®tability margins for the company. However, each retailer is expected to promote its own offerings and thus is not held to the standards of a search across competitors expected of an independent agent such as Bargain Finder. The third trust dimension involves the belief that the good intentions of the agents actually result in appropriate choices. This requirement was vividly illustrated in the failure of Blockbuster's Take 10 product. Providing poor suggestions may be worse than no suggestions at all.
How is trust developed? Urban, Sultan, and Qualls (1998) suggest that trust-building cues play a role. They consider the context of an online store with a shopping advisor, represented as a person with a pictorial image and name. Three classes of trust building cues are proposed for investigation. The ®rst group has to do with the site: issues such as privacy (e.g., no cookies), security of payments, and endorsements. The second deals with the shopping advisor and the advice given, for example: friendliness and warmth, unbiased data, and acknowledgment of weaknesses. The third group enters the process at the ful®llment stage and includes ease of ordering and payment, keeping promises, delivering on time, and avoidance of billing errors. Taken together this research offers a framework for studying the effect of cue characteristics on the building of customer con®dence and trust.
Additional research opportunities abound for testing ways to develop and preserve these dimensions of trust. First, there is the issue of the transparency of the agent's method. While collaborative ®ltering may produce good choice options, it is important to test various ways that the process can be explained to, and understood by consumers. Second, there is the related issue of controlÐthe perception that the consumer controls the agent rather than the reverse. This need for control suggests that it would be valuable for agents to ask screening criteria, and collect feedback on the desirably of suggested offerings. Third, there is the issue of framing the agent's recommendations. For example, should less valuable alternatives intentionally proceed the most desirable choice to increase its perceived value? Finally, research is needed into the patterns by which some types of businesses have natural advantages with respect to trust. For example, a customer may be more likely to trust a doctor or a bank with personal data than a supermarket (Hagel and Rayport 1997).
Notice that these research questions can be more easily answered in the context of electronic agents than for individuals or organizations. The processes by which electronic agents reinforce trust can be precisely de®ned in the programmed code, thereby minimiz-ing noise created by human and organizational factors. This greater control of the manipulations, combined with the ability to ef®ciently test strategies on many people simultaneously, leads to the prediction that theoretical and practical knowledge of trust will blossom through studies of electronic agents.
Note
1. This working paper is based on the ``Agents and Environments'' session of the 1998 HEC Invitational Choice Symposium. Please refrain from quoting or reproducing this document in whole or in part without the consent of the authors. Suggestions or feedback should be forwarded to the corresponding author: Patricia M. West, Associate Professor of Marketing, Fisher College of Business, The Ohio State University, 2100 Neil Ave., Room 544, Columbus, OH, 43210; e-mail: [email protected].
References
Ackoff, Russell L. (1967), ``Management Misinformation Systems,'' Management Science, 14(4), B147±B156. Alba, Joseph, John Lynch, Barton Weitz, Chris Janiszewski, Richard Lutz, Alan Sawyer, and Stacy Wood (1997), ``Interactive Home Shopping: Consumer, Retailer, and Manufacturer Incentives to Participate in Electronic Marketplaces,'' Journal of Marketing, 61 (July), 38±53.
Ariely, Dan (1998), ``Controlling the Information Flow: On the Role of Interactivity in Consumers' Decision Making and Preferences,'' working paper at the Sloan School of Management, MIT.
Avery, Christopher and Richard Zeckhauser (1997), ``Recommender Systems for Evaluating Computer Messages,'' Communications of the Association for Computing Machinery, 40(3), 88±90.
Bakos, J. Yannis (1997), ``Reducing Buyer Search Costs: Implications for Electronic Marketplaces,'' Management Science, 43 (12), 1676±1692.
Balzer, William K., Michael E. Doherty, and Raymond O'Connor, Jr. (1989), ``Effects of Cognitive Feedback on Performance,'' Psychological Bulletin, 106 (November), 410±433.
Bettman, James R., Mary Frances Luce, and John W. Payne (1998), ``Constructive Consumer Choice Processes,'' Journal of Consumer Research, 25 (December), 187±217.
Bettman, James R., Eric J. Johnson, and John W. Payne (1990), ``A Componential Analysis of Cognitive Effort in Choice,'' Organizational Behavior and Human Decision Processes, 45, 111±139.
Chase, William G. and Herbert A. Simon (1973), ``Perception in Chess,'' Cognitive Psychology, 4 (January), 55±81. Fischhoff, Baruch (1991), ``Value Elicitation: Is There Anything in There?'' American Psychologist, 46 (August),
835±847.
Fitzsimmons, Gavan J. (1996), ``Out-of-Stocks and Consumer Satisfaction,'' Unpublished doctoral dissertation, Columbia University, New York, NY.
Fitzsimmons, Gavan J., Eric A. Greenleaf, and Donald R. Lehmann (1997), ``Consumer Satisfaction with Both Product and Decision: Implications for the Supply Chain,'' UCLA working paper, 110 Westwood Plaza, Los Angeles, CA 90095.
Gershoff, Andrew and Patricia M. West (1998), ``Using a Community of Knowledge to Build Intelligent Agents,'' Marketing Letters, 9 (January), 79±91.
Gilmore, James H. and B. Joseph Pine II (1997), ``The Four Faces of Mass Customization,'' Harvard Business Review (January-February), 91±101.
Green, Paul E. A., M. Krieger, and M. K. Agarwal (1991), ``Adaptive Conjoint Analysis: Some Caveats and Suggestions,'' Journal of Marketing Research, Vol. 28, 215±222.
Greenleaf, Eric A. and Donald R. Lehmann (1995), ``Reasons for Substantial Delay in Consumer Decision Making,'' Journal of Consumer Research, 22 (2), 186±199
Hagel, John III and Jeffrey F. Rayport (1997), ``The Coming Battle for Customer Information,'' Harvard Business Review, (January-February), 53±65.
Huffman, Cynthia and Barbara. E. Kahn (1998), ``Variety for Sale: Mass Customization or Mass Confusion?'' Journal of Retailing, Winter, Vol. 74, # 2.
Kahn, Barbara E. (1995), ``Consumer Variety-Seeking Among Goods and Services: An Integrative Review,'' Journal of Retailing and Consumer Services, 2:3, 139±48.
Kahn, B. E. and Cynthia Huffman (1998), ``Mass Customization,'' The Handbook of Technology Management, Ed: R. C. Dorf, CRC Press, Inc.
Kahneman, Daniel and Snell, Jackie (1990), ``Predicting Utility,'' In R.M. Hogarth (ed.) Insights in Decision Making: A Tribute to Hillel J. Einhorn. Chicago: University of Chicago Press.
Keeney, Ralph L. (1992), Value-Focused Thinking: A Path to Creative Decisionmaking, Harvard Press, Cambridge MA.
Kleinmuntz, Donald N. and David A. Schkade (1993), ``Information Displays in Decision Making.'' Psycholo-gical Science, 4, 221±227.
Little, Roderick J. A and Donald B. Rubin (1987), Statistical Analysis with Missing Data, Wiley, New York. Loewenstein, George and David A. Schkade (1999), ``Wouldn't It Be Nice? Predicting Tastes and Feelings,'' In D.
Kahneman, E. Diener & N. Schwartz (Eds.), Well Being: The Foundations of Hedonic Psychology. Russell Sage Foundation.
Lohse, Jerry and Eric J. Johnson (1999), ``The In¯uence of Rearranging Product Information in a Matrix Display on Consumer Choice,'' The Wharton Forum for Electronic Commerce, working paper at the Wharton School, University of Pennsylvania.
Lynch, John G. and Dan Ariely (1998), ``Interactive Home Shopping: Effects of Search Costs for Price and Quality Information on Consumer Price Sensitivity, Satisfaction With Merchandise, and Retention,'' working paper at the Fuqua School of Business.
Maes, Pattie (1995), ``Intelligent Software,'' Scienti®c American, Vol. 273 (September), 84±86.
Malhotra, Naresk K. (1982), ``Information Load and Consumer Decision Making,'' Journal of Consumer Research, 8 (March), 419±430.
Mandel, Naomi and Eric Johnson (1998), ``Constructing Preferences Online: Can Web Pages Change What You Want?'' Working Paper, University of Pennsylvania.
McAlister, Leigh and Edgar A. Pessemier (1982), ``Variety-Seeking Behavior: An Interdisciplinary review,'' Journal of Consumer Research, 9 (December), 311±22.
Menon, Satya and Barbara E. Kahn (1998), ``Cross-Category Effects of Stimulation on the Shopping Experience: An Application to Internet Shopping,'' The Wharton School, University of Pennsylvania, Department of Marketing, Working Paper.
Payne, John, James R. Bettman, and Eric J. Johnson (1993), The Adaptive Decision Maker, Cambridge University Press; New York, NY.
Payne, John, James R. Bettman, and David A. Schkade (forthcoming), ``Measuring Constructed Preferences: Toward a Building Code,'' Journal of Risk and Uncertainty.
Pazgal, Amit and Nir Vulcan (1998), ``Have Your Agent Call Mine: Software Agents, the Internet and Marketing,'' in Donna Hoffman and John Little (Eds.), Collected Working Papers, Marketing Science and the Internet, INFORMS Mini-Conference, MIT Sloan School.
Reeves, Byron and Clifford Nass (1998), The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places, Cambridge University Press, New York, NY.
Russo, J. E. (1977), ``The Value of Unit Price Information,'' Journal of Marketing Research, 14, 193±201. Solomon, Michael R. (1986), ``The Missing Link: Surrogate Consumers in the Marketing Chain,'' Journal of
Marketing, 50, 208±218.
Urban, Glen (1998) ``Trust Based Marketing on the Web,'' in Donna Hoffman and John Little (Eds.), Collected. Working Paper, Marketing Science and the Internet, INFORMS Mini-Conference, MIT Sloan School. Urban, Glen L., Fareena Sultan, and William Qualls (1998), ``Trust Based Marketing on the Internet,'' M.I.T.
Sloan School Working Paper WP 4035±98.
West, Patricia M., Christina L. Brown, and Stephen J. Hoch (1996), ``Consumption Vocabulary and Preference Formation,'' Journal of Consumer Research, 22 (September), 120±135.
Wainer, H. and R. J. Mislevy (1990), ``Computerized Adaptive Testing: A Primer,'' Lawrence Erlbaum Associates, Hillsdale, NJ.
Widing, Robert and W. Wayne Talarzyk (1993), ``Electronic Information Systems for Consumers: An Evaluation of Computer-Assisted Formats in Multiple Decision Environments,'' Journal of Marketing Research, 30 (May), 125±141.
